Data-augmented DD-SIMCA models for the authentication of aged cachaça using digital images

The authentication of aged cachaças poses a challenge for quality control, which is often limited by the high cost and difficulty of obtaining a statistically relevant number of reference samples. Single-class models, such as DD-SIMCA, require representative training sets to ensure high sensitivity and specificity. In this study, we propose using synthetic data augmentation techniques to expand the chemical space of real samples and improve the sensitivity, specificity, and efficiency of color-histogram-based authentication models. The dataset consisted of color histograms (Full, GS, GS+HSV, GS+RGB, HSV, RGB, and RG+BHSV) extracted from digital images of aged cachaças from Amburana and Carvalho (20 samples per class). Two data augmentation strategies were investigated: PCA sampling, via reconstruction using simulated scores, and SMOTE, based on nearest-neighbor interpolation, compared with the conventional scenario without sample expansion. Data partitioning (75% training / 25% testing) was performed using the Kennard-Stone algorithm. DD-SIMCA models (α = 0.05; robust estimation) were constructed and validated against real samples and samples adulterated with wood extracts (20 samples per group). The integration of data augmentation GS+RGB dataset affected the DD-SIMCA architecture, in which the conventional model achieved 100% sensitivity, 97.5% specificity, and 98.7% efficiency with 8 Principal Components (PCs). At a 1:1 augmentation ratio, both PCA-sampling and SMOTE maintained identical external prediction performance (100% sensitivity, 97.5% specificity), but induced a subtle drop in training sensitivity to 95%. This indicates that synthetic samples compressed and refined the critical decision thresholds, making the statistical envelope more stringent and increasing the rejection of true boundary samples during training without compromising external generalization. A clear divergence emerged in complexity and signal morphology, with PCA sampling yielding a more parsimonious architecture with 7 PCs. This occurred because PCA sampling operates under a global multivariate normal distribution in the latent space, generating smoothed profiles shifted toward the class center. Conversely, SMOTE required 8 PCs because it simulated the narrow, sharp peaks of the color histograms with superior fidelity, especially within the complex variations of the blue channel. By performing local linear interpolations between immediate geometric neighbors, SMOTE preserved the sharp morphology and spatial colinearity of the original optical signals. Therefore, while both methods successfully stabilize one-class boundaries against smart wood-extract fraud, SMOTE offers a more realistic representation of discrete camera sensor outputs. In contrast, PCA sampling provides a more compact mathematical description of the target space. Data augmentation effectively refines and tightens the critical decision thresholds of one-class DD-SIMCA models, enhancing boundary stringency for aged cachaça authentication using digital images.

Acknowledgments: The authors thank CAPES for the financial support and the research scholarship.